A modular and reproducible pipeline for univariate time series forecasting using ARIMA, LSTM, and Transformer models. This repository is designed for plug-and-play usage on electricity demand data, but can be easily adapted to other time series datasets.
This repo includes the following autoregressive forecasting models:
- ARIMA: Classical linear model with seasonal extensions.
- LSTM: Recurrent neural network capable of learning temporal dependencies.
- Transformer: A causal, decoder-style transformer model with masked self-attention and positional encoding, tailored for univariate forecasting.
Experiments are run on UK national electricity demand data (NESO, 2024), with all models predicting one step ahead over a 7-day test window (336 half-hourly steps).
